newmark-agent 0.5.4 → 0.5.7

package/dist/conversation-utility-host.bundle.cjs

@@ -328670,7 +328670,7 @@ function providerStreamTimeoutError(timeoutMs) {
   error.message = `Stream read timeout after ${timeoutMs}ms`;
   return error;
 }
-async function readProviderStreamChunk(reader, signal, timeoutMs = 3e4) {
+async function readProviderStreamChunk(reader, signal, timeoutMs = 0) {
   if (signal.aborted) throw providerAbortError(signal);
   let timer;
   let onAbort;
@@ -328678,9 +328678,9 @@ async function readProviderStreamChunk(reader, signal, timeoutMs = 3e4) {
     onAbort = () => reject(providerAbortError(signal));
     signal.addEventListener("abort", onAbort, { once: true });
   });
-  const timeoutPromise = new Promise((_3, reject) => {
+  const timeoutPromise = timeoutMs > 0 ? new Promise((_3, reject) => {
     timer = setTimeout(() => reject(providerStreamTimeoutError(timeoutMs)), timeoutMs);
-  });
+  }) : new Promise(() => void 0);
   try {
     return await Promise.race([reader.read(), abortPromise, timeoutPromise]);
   } catch (error) {
@@ -340857,7 +340857,13 @@ async function runAgentKernel(agent) {
       agent.recordWorkStatus("Final visual fallback used: local mini OCR plus conservative text correction.");
     }
     if (modelBeforeKernelRun && modelBeforeKernelRun !== agent.model && !tokens.some((t3) => t3.text?.includes("[Model fallback]"))) {
-      tokens.unshift({ type: "text", text: `[Model fallback] ${modelBeforeKernelRun} unavailable; switched to ${agent.model}.` });
+      const notice = `[Model fallback] ${modelBeforeKernelRun} unavailable; switched to ${agent.model}.`;
+      tokens.unshift({ type: "text", text: notice });
+      agent.emitWorkEvent({
+        type: "status",
+        content: notice,
+        fallback: { from: modelBeforeKernelRun, to: agent.model, providerId: agent.activeDeployment()?.providerId }
+      });
     }
     let emptyResponseRetries = 0;
     while (providerTurnIsEmpty(lastTurn) && emptyResponseRetries < 2) {
@@ -340878,6 +340884,11 @@ async function runAgentKernel(agent) {
       const notice = routeTransitionNotice(agent, previous);
       tokens.push({ type: "text", text: notice });
       agent.recordWorkStatus(notice);
+      agent.emitWorkEvent({
+        type: "status",
+        content: notice,
+        fallback: { from: previous, to: agent.model, providerId: agent.activeDeployment()?.providerId }
+      });
       kernel2.state.model = toKernelModel(agent);
       const fallbackToolSurface = refreshToolSurface(true);
       kernel2.state.systemPrompt = [agent.buildSystemPrompt(), fallbackToolSurface.systemPromptNotice].filter(Boolean).join("\n\n");
@@ -343302,14 +343313,16 @@ var AutoRouter = class {
         retryDelayMs
       });
     }
+    if (failure.allowModelFallback === false) return attempts.slice(0, remainingAttempts);
     const selection = decision.requestedSelection;
     if (selection.kind === "fixed") return attempts.slice(0, remainingAttempts);
     const scope = selection.kind === "auto" ? selection.scope : { kind: "provider", providerId: current.providerId };
     const subset = selection.kind === "auto" ? selection.subset : void 0;
     const currentGroup = current.logicalModelGroupId;
+    const currentProviderId = current.providerId;
     const now2 = this.now();
     const attemptedDeployments = decision.attempts.map((attempt) => attempt.deployment);
-    const eligible = candidates.filter((candidate) => candidate.enabled && inScope(candidate.deployment, scope) && inSubset(candidate.deployment, subset) && !sameDeployment(candidate.deployment, current) && !attemptedDeployments.some((attempted) => sameDeployment(candidate.deployment, attempted)) && validationEligible(candidate, now2).length === 0 && this.passedInitialHardFilters(decision, candidate) && this.circuitState(candidate.deployment, now2, false) !== "open");
+    const eligible = candidates.filter((candidate) => candidate.enabled && candidate.deployment.providerId === currentProviderId && inScope(candidate.deployment, scope) && inSubset(candidate.deployment, subset) && !sameDeployment(candidate.deployment, current) && !attemptedDeployments.some((attempted) => sameDeployment(candidate.deployment, attempted)) && validationEligible(candidate, now2).length === 0 && this.passedInitialHardFilters(decision, candidate) && this.circuitState(candidate.deployment, now2, false) !== "open");
     const equivalent = currentGroup ? eligible.find((candidate) => candidate.deployment.logicalModelGroupId === currentGroup && !candidate.fallbackOnly) : void 0;
     const fallback = eligible.find((candidate) => candidate.fallbackOnly);
     const rankedAlternates = decision.rankedCandidates.map((ranked) => eligible.find((candidate) => sameDeployment(candidate.deployment, ranked.deployment))).filter((candidate) => !!candidate && candidate !== equivalent && candidate !== fallback && !candidate.fallbackOnly);
@@ -346747,7 +346760,8 @@ ${String(event.toolArgs || "")}`;
       sequence,
       status: input.status,
       guide: !isToolEvent && input.guide ? this.normalizeGuideReceipt(input.guide) : void 0,
-      displayImage: isToolEvent ? this.hydrateDisplayImage(input.displayImage) : void 0
+      displayImage: isToolEvent ? this.hydrateDisplayImage(input.displayImage) : void 0,
+      fallback: input.fallback
     };
     if (activeRun && this.isPersistablePublicWorkEvent(event)) {
       activeRun.sequence = Number(sequence || activeRun.sequence + 1);
@@ -349666,9 +349680,15 @@ ${summary}`, segment, "local-summarize", true);
   }
   compressionBuildBlockStart(messages) {
     const activeRunId = this.currentWorkRunId();
-    if (!activeRunId) return 0;
-    const index = messages.findIndex((message) => String(message.run_id || message.runId || "") === activeRunId);
-    return index >= 0 ? index : 0;
+    if (activeRunId) {
+      const index = messages.findIndex((message) => String(message.run_id || message.runId || "") === activeRunId);
+      if (index >= 0) return index;
+    }
+    let lastRunBoundary = -1;
+    for (let index = 0; index < messages.length; index += 1) {
+      if (String(messages[index]?.run_id || messages[index]?.runId || "")) lastRunBoundary = index + 1;
+    }
+    return lastRunBoundary >= 0 ? lastRunBoundary : messages.length;
   }
   recentContextSuffix(messages, maxMessages, tokenBudget) {
     if (!messages.length) return [];
@@ -350597,17 +350617,12 @@ ${msg.content}
         currentAttempt.sideEffectBoundary = this.routeSideEffectCommitted;
         if (this.routeAttemptStartedAt > 0) currentAttempt.durationMs = Math.max(0, Date.now() - this.routeAttemptStartedAt);
       }
-      if (!fallbackEnabled) {
-        this.lastRouteDecision.finalStatus = "failed";
-        this.routeAttemptStartedAt = 0;
-        this.persistRouteDecision(this.lastRouteDecision);
-        return null;
-      }
       if (!this.pendingAutoAttempts.length) {
         this.pendingAutoAttempts = this.autoRouter.planAttempts(this.lastRouteDecision, this.autoRouteCandidates(), {
           error: failure,
           streamCommitted: this.routeStreamCommitted,
-          sideEffectCommitted: this.routeSideEffectCommitted
+          sideEffectCommitted: this.routeSideEffectCommitted,
+          allowModelFallback: fallbackEnabled
         });
       }
       const next2 = this.pendingAutoAttempts.shift();
@@ -350642,7 +350657,6 @@ ${msg.content}
     if (!next?.name) return null;
     this.model = next.name;
     this.fixedDeployment = this.deploymentRef(next);
-    if (next.provider) this.config.set("models", "auto_switch_anchor_provider", next.provider);
     return current;
   }
   isLlmErrorText(text) {
@@ -350650,12 +350664,10 @@ ${msg.content}
   }
   scopedSwitchModels(currentModelName) {
     const all = this.config.allModels();
-    if (this.config.autoSwitchScope() !== "provider") return all;
-    const current = currentModelName === "auto" ? void 0 : this.config.findModel(currentModelName);
-    const providerId = current?.provider_id || this.config.autoSwitchAnchorProvider() || this.config.findModel(this.config.getStr("models", "default_model"))?.provider_id || all[0]?.provider_id || "";
-    if (!providerId) return all;
-    const provider = this.config.findProvider(providerId);
-    return all.filter((m2) => m2.provider_id === (provider?.id || providerId));
+    const current = currentModelName === "auto" ? this.activeModelConfig() : this.config.findModel(currentModelName);
+    const providerId = current?.provider_id || this.fixedDeployment?.providerId || (currentModelName !== "auto" ? this.activeDeployment()?.providerId : void 0);
+    if (!providerId) return [];
+    return all.filter((m2) => m2.provider_id === providerId);
   }
   async validateModels(selectedNames, options = {}) {
     if (this.modelValidationPromise) return this.modelValidationPromise;
@@ -352037,7 +352049,10 @@ ${items.map((item) => this.formatAutomation(item)).join("\n")}`;
     if (!sa) return "[Subagent] Not found.";
     const requestedModel = this.normalizeSubagentModelSelection(sa.model && sa.model !== "default" ? sa.model : this.modelSelectionValue() || this.model);
     const requestedDeployment = parseDeploymentSelectionValue2(requestedModel);
-    const assignedModel = requestedModel === "auto" ? this.activeModelConfig() : requestedDeployment ? this.config.findDeployment(requestedDeployment) : this.config.findModel(requestedModel);
+    let assignedModel = requestedModel === "auto" ? this.activeModelConfig() : requestedDeployment ? this.config.findDeployment(requestedDeployment) : this.config.findModel(requestedModel);
+    if (!assignedModel && !requestedDeployment && requestedModel !== "auto") {
+      assignedModel = this.activeModelConfig();
+    }
     const model = assignedModel?.name || (requestedModel === "auto" ? this.activeModelName() : requestedModel);
     const activeModel = this.activeModelConfig();
     const activeProvider = this.engineModel();
@@ -353466,6 +353481,32 @@ var ConversationKernel = class {
       followUp: queued?.followUp.slice() || []
     };
   }
+  /**
+   * dev-0.5.6: 排队 followUp 消息 drain 时转成普通用户消息载荷。
+   *
+   * 入队时（enqueueSameSession）为去重/编辑保留了 clientMessageId，并把
+   * runId 固定为入队时运行的 runId；若原样传给 process，写入 chatMessages
+   * 的用户消息会携带旧 runId + clientMessageId，渲染端（PC renderTranscript /
+   * 移动端 projectRemoteConversationItems）会把「clientMessageId + runId 命中
+   * workRuns」的消息误判为 guide 消息而跳过气泡，导致排队自动发送的消息不
+   * 显示在用户输入时间线。清除这两个字段后 process 走普通用户消息分支
+   * （chatMessages.push，mode 取 visibleMode，runId 取当前 run）。
+   */
+  drainQueuedFollowUpMessage(message) {
+    if (typeof message === "string") return message;
+    const text = String(message.text || "");
+    const images = message.images;
+    const attachments = message.attachments;
+    const visible = message.visibleUserInput;
+    const visibleMode = message.visibleMode;
+    return {
+      text,
+      ...images?.length ? { images } : {},
+      ...attachments?.length ? { attachments } : {},
+      ...visible ? { visibleUserInput: visible } : {},
+      ...visibleMode ? { visibleMode } : {}
+    };
+  }
   queueItems(target) {
     const runtime = this.findRuntime(target);
     if (!runtime) return [];
@@ -354207,7 +354248,7 @@ ${batchText}`,
           };
           lastTokens = await this.runSingle(runtime, batchMessage, "steer");
         } else {
-          lastTokens = await this.runSingle(runtime, next.message, next.queueMode);
+          lastTokens = await this.runSingle(runtime, this.drainQueuedFollowUpMessage(next.message), next.queueMode);
         }
       }
       const rootMessage = runtime.runner.subagents.readRootInbox()[0];
@@ -354666,7 +354707,11 @@ ${text}`;
       type: "queue_update",
       content: "Conversation queue updated.",
       conversationId: runtime.id,
-      queue: this.queued(runtime.target)
+      queue: this.queued(runtime.target),
+      // Structured rows with stable kernel ids so every consumer (PC UI and
+      // the paired mobile client) can render/edit/delete the same items.
+      queueItems: this.queueItems(runtime.target),
+      queuePaused: runtime.queuePaused === true
     });
   }
   clearQueued(runtime) {

package/dist/core/agent.js

@@ -2298,6 +2298,7 @@ class Agent {
             status: input.status,
             guide: !isToolEvent && input.guide ? this.normalizeGuideReceipt(input.guide) : undefined,
             displayImage: isToolEvent ? this.hydrateDisplayImage(input.displayImage) : undefined,
+            fallback: input.fallback,
         };
         if (activeRun && this.isPersistablePublicWorkEvent(event)) {
             activeRun.sequence = Number(sequence || activeRun.sequence + 1);
@@ -5587,10 +5588,22 @@ class Agent {
     }
     compressionBuildBlockStart(messages) {
         const activeRunId = this.currentWorkRunId();
-        if (!activeRunId)
-            return 0;
-        const index = messages.findIndex(message => String(message.run_id || message.runId || '') === activeRunId);
-        return index >= 0 ? index : 0;
+        if (activeRunId) {
+            const index = messages.findIndex(message => String(message.run_id || message.runId || '') === activeRunId);
+            if (index >= 0)
+                return index;
+        }
+        // dev-0.5.6: 空闲对话（无活动 run）时不得把全部历史算进当前 Build Block，
+        // 否则上下文显示窗口的长期历史恒为 0。回退语义：
+        // - 存在带 run_id 的消息：boundary 取最后一个 run 的起点之后（该 run 及其
+        //   之前的历史属于长期历史，其后无归属的消息属于当前未命名区块）；
+        // - 完全没有 run_id：全部历史都是长期历史（不存在当前 Build Block）。
+        let lastRunBoundary = -1;
+        for (let index = 0; index < messages.length; index += 1) {
+            if (String(messages[index]?.run_id || messages[index]?.runId || ''))
+                lastRunBoundary = index + 1;
+        }
+        return lastRunBoundary >= 0 ? lastRunBoundary : messages.length;
     }
     recentContextSuffix(messages, maxMessages, tokenBudget) {
         if (!messages.length)
@@ -6651,17 +6664,12 @@ class Agent {
                 if (this.routeAttemptStartedAt > 0)
                     currentAttempt.durationMs = Math.max(0, Date.now() - this.routeAttemptStartedAt);
             }
-            if (!fallbackEnabled) {
-                this.lastRouteDecision.finalStatus = 'failed';
-                this.routeAttemptStartedAt = 0;
-                this.persistRouteDecision(this.lastRouteDecision);
-                return null;
-            }
             if (!this.pendingAutoAttempts.length) {
                 this.pendingAutoAttempts = this.autoRouter.planAttempts(this.lastRouteDecision, this.autoRouteCandidates(), {
                     error: failure,
                     streamCommitted: this.routeStreamCommitted,
                     sideEffectCommitted: this.routeSideEffectCommitted,
+                    allowModelFallback: fallbackEnabled,
                 });
             }
             const next = this.pendingAutoAttempts.shift();
@@ -6702,8 +6710,11 @@ class Agent {
             return null;
         this.model = next.name;
         this.fixedDeployment = this.deploymentRef(next);
-        if (next.provider)
-            this.config.set('models', 'auto_switch_anchor_provider', next.provider);
+        // Do NOT rewrite auto_switch_anchor_provider here. The anchor describes
+        // the user's Auto-routing scope; a fixed-model recovery must never
+        // repoint it (a single transient failure would otherwise silently move
+        // every future Auto route — and every future fallback pool — to the
+        // recovery provider, which is the cross-provider leak users observed).
         return current;
     }
     isLlmErrorText(text) {
@@ -6711,18 +6722,23 @@ class Agent {
     }
     scopedSwitchModels(currentModelName) {
         const all = this.config.allModels();
-        if (this.config.autoSwitchScope() !== 'provider')
-            return all;
-        const current = currentModelName === 'auto' ? undefined : this.config.findModel(currentModelName);
-        const providerId = current?.provider_id ||
-            this.config.autoSwitchAnchorProvider() ||
-            this.config.findModel(this.config.getStr('models', 'default_model'))?.provider_id ||
-            all[0]?.provider_id ||
-            '';
+        // Fixed-model recovery is always provider-local. A same-named model on a
+        // different provider is a different deployment and must never receive the
+        // failed provider's credentials or become an implicit fallback.
+        // Resolution order is strictly deployment-first: the pinned deployment,
+        // then the exact active deployment. The global auto anchor is deliberately
+        // NOT consulted here — it belongs to Auto routing only and can point at a
+        // different provider (for example after a previous cross-provider switch),
+        // which is exactly the same-name cross-provider leak this guard exists to
+        // prevent. When no provider can be established unambiguously, return an
+        // empty pool: failing the switch is safe, crossing providers is not.
+        const current = currentModelName === 'auto' ? this.activeModelConfig() : this.config.findModel(currentModelName);
+        const providerId = current?.provider_id
+            || this.fixedDeployment?.providerId
+            || (currentModelName !== 'auto' ? this.activeDeployment()?.providerId : undefined);
         if (!providerId)
-            return all;
-        const provider = this.config.findProvider(providerId);
-        return all.filter(m => m.provider_id === (provider?.id || providerId));
+            return [];
+        return all.filter(m => m.provider_id === providerId);
     }
     async validateModels(selectedNames, options = {}) {
         if (this.modelValidationPromise)
@@ -8224,9 +8240,17 @@ class Agent {
             return '[Subagent] Not found.';
         const requestedModel = this.normalizeSubagentModelSelection(sa.model && sa.model !== 'default' ? sa.model : this.modelSelectionValue() || this.model);
         const requestedDeployment = parseDeploymentSelectionValue(requestedModel);
-        const assignedModel = requestedModel === 'auto'
+        let assignedModel = requestedModel === 'auto'
             ? this.activeModelConfig()
             : (requestedDeployment ? this.config.findDeployment(requestedDeployment) : this.config.findModel(requestedModel));
+        // Imported presets (and stale records) may reference a bare model name
+        // that no longer exists in the configured catalog. Resolving by name must
+        // never guess a provider; instead the peer inherits the parent deployment
+        // so the job can still run. Qualified deployment values keep failing
+        // closed — they are exact references, not guesses.
+        if (!assignedModel && !requestedDeployment && requestedModel !== 'auto') {
+            assignedModel = this.activeModelConfig();
+        }
         const model = assignedModel?.name || (requestedModel === 'auto' ? this.activeModelName() : requestedModel);
         const activeModel = this.activeModelConfig();
         const activeProvider = this.engineModel();

package/dist/core/agentKernelRunner.js

@@ -440,7 +440,13 @@ async function runAgentKernel(agent) {
             agent.recordWorkStatus('Final visual fallback used: local mini OCR plus conservative text correction.');
         }
         if (modelBeforeKernelRun && modelBeforeKernelRun !== agent.model && !tokens.some(t => t.text?.includes('[Model fallback]'))) {
-            tokens.unshift({ type: 'text', text: `[Model fallback] ${modelBeforeKernelRun} unavailable; switched to ${agent.model}.` });
+            const notice = `[Model fallback] ${modelBeforeKernelRun} unavailable; switched to ${agent.model}.`;
+            tokens.unshift({ type: 'text', text: notice });
+            agent.emitWorkEvent({
+                type: 'status',
+                content: notice,
+                fallback: { from: modelBeforeKernelRun, to: agent.model, providerId: agent.activeDeployment()?.providerId },
+            });
         }
         let emptyResponseRetries = 0;
         while (providerTurnIsEmpty(lastTurn) && emptyResponseRetries < 2) {
@@ -462,6 +468,11 @@ async function runAgentKernel(agent) {
             const notice = routeTransitionNotice(agent, previous);
             tokens.push({ type: 'text', text: notice });
             agent.recordWorkStatus(notice);
+            agent.emitWorkEvent({
+                type: 'status',
+                content: notice,
+                fallback: { from: previous, to: agent.model, providerId: agent.activeDeployment()?.providerId },
+            });
             kernel.state.model = toKernelModel(agent);
             const fallbackToolSurface = refreshToolSurface(true);
             kernel.state.systemPrompt = [agent.buildSystemPrompt(), fallbackToolSurface.systemPromptNotice].filter(Boolean).join('\n\n');

package/dist/core/autoRouter.d.ts

@@ -153,6 +153,7 @@ export declare class AutoRouter {
         error: RouteFailure;
         streamCommitted: boolean;
         sideEffectCommitted: boolean;
+        allowModelFallback?: boolean;
     }): PlannedRouteAttempt[];
     recordEndpointFailure(deployment: DeploymentRef, failure: RouteFailureType | RouteFailure): void;
     recordEndpointSuccess(deployment: DeploymentRef, latencyMs?: number, throughput?: number): void;

package/dist/core/autoRouter.js

@@ -290,15 +290,24 @@ class AutoRouter {
                 retryDelayMs,
             });
         }
+        if (failure.allowModelFallback === false)
+            return attempts.slice(0, remainingAttempts);
         const selection = decision.requestedSelection;
         if (selection.kind === 'fixed')
             return attempts.slice(0, remainingAttempts);
         const scope = selection.kind === 'auto' ? selection.scope : { kind: 'provider', providerId: current.providerId };
         const subset = selection.kind === 'auto' ? selection.subset : undefined;
         const currentGroup = current.logicalModelGroupId;
+        const currentProviderId = current.providerId;
         const now = this.now();
         const attemptedDeployments = decision.attempts.map(attempt => attempt.deployment);
+        // Fallback is a recovery operation, not a new global routing decision.
+        // It must never cross the provider boundary, even when the original Auto
+        // selection has global scope. This prevents equal model ids from silently
+        // switching credentials/endpoints (for example provider A/model X to
+        // provider B/model X).
         const eligible = candidates.filter(candidate => candidate.enabled
+            && candidate.deployment.providerId === currentProviderId
             && inScope(candidate.deployment, scope)
             && inSubset(candidate.deployment, subset)
             && !sameDeployment(candidate.deployment, current)

package/dist/core/conversationKernel.d.ts

@@ -194,6 +194,18 @@ export declare class ConversationKernel {
         steering: string[];
         followUp: string[];
     };
+    /**
+     * dev-0.5.6: 排队 followUp 消息 drain 时转成普通用户消息载荷。
+     *
+     * 入队时（enqueueSameSession）为去重/编辑保留了 clientMessageId，并把
+     * runId 固定为入队时运行的 runId；若原样传给 process，写入 chatMessages
+     * 的用户消息会携带旧 runId + clientMessageId，渲染端（PC renderTranscript /
+     * 移动端 projectRemoteConversationItems）会把「clientMessageId + runId 命中
+     * workRuns」的消息误判为 guide 消息而跳过气泡，导致排队自动发送的消息不
+     * 显示在用户输入时间线。清除这两个字段后 process 走普通用户消息分支
+     * （chatMessages.push，mode 取 visibleMode，runId 取当前 run）。
+     */
+    private drainQueuedFollowUpMessage;
     queueItems(target: ConversationTargetInput): ConversationQueueItemSnapshot[];
     enqueueNext(target: ConversationTargetInput, input: {
         id: string;

package/dist/core/conversationKernel.js

@@ -60,6 +60,33 @@ class ConversationKernel {
             followUp: queued?.followUp.slice() || [],
         };
     }
+    /**
+     * dev-0.5.6: 排队 followUp 消息 drain 时转成普通用户消息载荷。
+     *
+     * 入队时（enqueueSameSession）为去重/编辑保留了 clientMessageId，并把
+     * runId 固定为入队时运行的 runId；若原样传给 process，写入 chatMessages
+     * 的用户消息会携带旧 runId + clientMessageId，渲染端（PC renderTranscript /
+     * 移动端 projectRemoteConversationItems）会把「clientMessageId + runId 命中
+     * workRuns」的消息误判为 guide 消息而跳过气泡，导致排队自动发送的消息不
+     * 显示在用户输入时间线。清除这两个字段后 process 走普通用户消息分支
+     * （chatMessages.push，mode 取 visibleMode，runId 取当前 run）。
+     */
+    drainQueuedFollowUpMessage(message) {
+        if (typeof message === 'string')
+            return message;
+        const text = String(message.text || '');
+        const images = message.images;
+        const attachments = message.attachments;
+        const visible = message.visibleUserInput;
+        const visibleMode = message.visibleMode;
+        return {
+            text,
+            ...(images?.length ? { images } : {}),
+            ...(attachments?.length ? { attachments } : {}),
+            ...(visible ? { visibleUserInput: visible } : {}),
+            ...(visibleMode ? { visibleMode } : {}),
+        };
+    }
     queueItems(target) {
         const runtime = this.findRuntime(target);
         if (!runtime)
@@ -888,7 +915,7 @@ class ConversationKernel {
                     lastTokens = await this.runSingle(runtime, batchMessage, 'steer');
                 }
                 else {
-                    lastTokens = await this.runSingle(runtime, next.message, next.queueMode);
+                    lastTokens = await this.runSingle(runtime, this.drainQueuedFollowUpMessage(next.message), next.queueMode);
                 }
             }
             const rootMessage = runtime.runner.subagents.readRootInbox()[0];
@@ -1414,6 +1441,10 @@ class ConversationKernel {
             content: 'Conversation queue updated.',
             conversationId: runtime.id,
             queue: this.queued(runtime.target),
+            // Structured rows with stable kernel ids so every consumer (PC UI and
+            // the paired mobile client) can render/edit/delete the same items.
+            queueItems: this.queueItems(runtime.target),
+            queuePaused: runtime.queuePaused === true,
         });
     }
     clearQueued(runtime) {

package/dist/core/electronUtilityRuntimePool.js

@@ -601,8 +601,13 @@ class ElectronUtilityRuntimePool {
         return this.accessSequence;
     }
     maxResidentRuntimes() {
-        const configured = Number(this.options.maxResidentRuntimes ?? 2);
-        return Number.isFinite(configured) ? Math.max(1, Math.floor(configured)) : 2;
+        // The pool hosts one utility runtime per active conversation target.
+        // A small default (2) silently blocked users from running more than two
+        // conversations at once with "utility runtime pool capacity reached".
+        // Idle runtimes are still evicted on LRU after the idle TTL, so a higher
+        // default bounds memory by activity, not by an arbitrary conversation cap.
+        const configured = Number(this.options.maxResidentRuntimes ?? 8);
+        return Number.isFinite(configured) ? Math.max(1, Math.floor(configured)) : 8;
     }
     async serializeCapacity(operation) {
         const previous = this.capacityTail;

package/dist/core/types.d.ts

@@ -112,6 +112,17 @@ export interface AgentWorkEvent {
         steering: string[];
         followUp: string[];
     };
+    /** Structured queue rows with stable kernel ids (mobile/PC queue unification export). */
+    queueItems?: Array<{
+        id: string;
+        text: string;
+        queueMode: string;
+        requestedMode?: string;
+        goalObjective?: string;
+        runId?: string;
+        createdAt: string;
+    }>;
+    queuePaused?: boolean;
     workspaceId?: string;
     workspaceKey?: string;
     runtimeKey?: string;
@@ -124,6 +135,16 @@ export interface AgentWorkEvent {
     status?: GuideReceiptStatus | ConversationWorkRunStatus | 'stopping' | 'force_restarting';
     guide?: GuideReceipt;
     displayImage?: DisplayImageAttachment;
+    /**
+     * 结构化模型回退信号：from 为回退前的模型名，to 为实际使用的模型名。
+     * 前端据此把输入框下方的模型选择区同步为实际生效的模型，而不是隐藏的
+     * 参数回退。
+     */
+    fallback?: {
+        from: string;
+        to: string;
+        providerId?: string;
+    };
 }
 export interface ConversationWorkRun {
     runId: string;

package/dist/core/wslAgentRuntimePool.js

@@ -569,8 +569,12 @@ class WslAgentRuntimePool {
         return this.accessSequence;
     }
     maxResidentRuntimes() {
-        const configured = Number(this.options.maxResidentRuntimes ?? 2);
-        return Number.isFinite(configured) ? Math.max(1, Math.floor(configured)) : 2;
+        // Same rationale as the Electron utility pool: one runtime per active
+        // conversation target. The previous default of 2 blocked parallel
+        // conversations with a capacity error; idle LRU eviction still bounds
+        // resident memory.
+        const configured = Number(this.options.maxResidentRuntimes ?? 8);
+        return Number.isFinite(configured) ? Math.max(1, Math.floor(configured)) : 8;
     }
     async serializeCapacity(operation) {
         const previous = this.capacityTail;

package/dist/main.js

@@ -3020,6 +3020,10 @@ else {
                     activePromptLeases.set(promptLeaseKey, (activePromptLeases.get(promptLeaseKey) || 0) + 1);
                     activePromptWorkspaces.set(promptLeaseWorkspaceKey, (activePromptWorkspaces.get(promptLeaseWorkspaceKey) || 0) + 1);
                     const targetConversation = target.conversationId;
+                    // 发送命令时锁定输入框选择的模型：接受命令后的整个运行过程都以该
+                    // 模型为准（唯一例外是显式的不可用回退，且回退会以结构化事件同步
+                    // 到前端输入框下方的选择区，而不是隐藏的参数回退）。
+                    const requestedModel = agent.model;
                     const options = {
                         mode: agent.mode,
                         model: agent.ensureUsableModelSelection(),
@@ -3027,6 +3031,21 @@ else {
                         inputMode: agent.inputMode,
                         engine: agent.engine,
                     };
+                    if (options.model && requestedModel && options.model !== requestedModel) {
+                        const usableConfig = agent.activeModelConfig();
+                        broadcastAgentWorkEvent({
+                            id: `model-fallback-${Date.now()}-${Math.random().toString(16).slice(2)}`,
+                            conversationId: target.conversationId,
+                            type: 'status',
+                            content: `[Model fallback] ${requestedModel} unavailable; switched to ${options.model}.`,
+                            mode: agent.modeName(),
+                            model: options.model,
+                            timestamp: new Date().toISOString(),
+                            workspaceId: target.workspaceId,
+                            workspaceKey: target.workspaceKey,
+                            fallback: { from: requestedModel, to: options.model, providerId: usableConfig?.provider_id || agent.activeDeployment()?.providerId },
+                        });
+                    }
                     const queueMode = agent.inputMode === 'guide' ? 'steer' : 'followUp';
                     let result;
                     if (wslBackendEnabled()) {
@@ -5206,6 +5225,21 @@ else {
                 // make an ordinary minimize click disappear from the taskbar.
                 win?.minimize();
             });
+            electron_1.ipcMain.handle('glass:captureBackdrop', async (event, requestedSize) => {
+                const image = await event.sender.capturePage();
+                const sourceSize = image.getSize();
+                const width = Math.max(1, Math.min(sourceSize.width, Math.round(Number(requestedSize?.width) || sourceSize.width)));
+                const height = Math.max(1, Math.min(sourceSize.height, Math.round(Number(requestedSize?.height) || sourceSize.height)));
+                const resized = sourceSize.width === width && sourceSize.height === height
+                    ? image
+                    : image.resize({ width, height, quality: 'good' });
+                return {
+                    bytes: resized.toJPEG(82),
+                    mimeType: 'image/jpeg',
+                    width,
+                    height,
+                };
+            });
             electron_1.ipcMain.handle('app:maximize', () => {
                 const win = electron_1.BrowserWindow.getFocusedWindow() || mainWindow;
                 if (win?.isMaximized())

package/dist/preload.js

@@ -6,6 +6,7 @@ contextBridge.exposeInMainWorld('api', {
     startupWaitForBackend: () => ipcRenderer.invoke('startup:waitForBackend'),
     startupUiReady: (payload) => ipcRenderer.invoke('startup:uiReady', payload),
     startupUiFailed: (payload) => ipcRenderer.invoke('startup:uiFailed', payload),
+    captureLiquidBackdrop: (size) => ipcRenderer.invoke('glass:captureBackdrop', size),
     onStartupStatus: (callback) => {
         ipcRenderer.on('startup:status', (_event, payload) => callback(payload));
     },

package/dist/providers/provider-events.d.ts

@@ -17,6 +17,12 @@ export declare function providerStreamTimeoutError(timeoutMs: number): Error;
  * Read one SSE chunk with both user cancellation and an inactivity deadline.
  * Cancelling the reader is important: rejecting the race alone leaves the
  * provider socket alive and lets later requests accumulate behind it.
+ *
+ * timeoutMs defaults to 0 (no stream idle deadline), matching the request-
+ * level DEFAULT_PROVIDER_REQUEST_TIMEOUT_MS = 0 and the Android client's
+ * readTimeout(0) / SSE_IDLE_TIMEOUT_MS = 0L. A caller that still wants an
+ * inactivity cap passes an explicit positive value (the recovery verify
+ * passes 50ms to prove reader cancellation).
  */
 export declare function readProviderStreamChunk(reader: ReadableStreamDefaultReader<Uint8Array>, signal: AbortSignal, timeoutMs?: number): Promise<ReadableStreamReadResult<Uint8Array>>;
 export declare function parseProviderSse(raw: string): Array<{